About us

2 — Objectives

AI4CCAM aims at developing an open environment for Trustworthy AI in CCAM operations, including:

• Simulated and synthetic urban-traffic scenarios involving VRUs and an interoperable digital framework to manage and multiply these scenarios in conformance with Trustworthy AI requirements and security measures
• Documentation about the risks related to the usage of AI in CAVs, an online participatory space for assessing and fostering the adoption of Trustworthy AI in CCAM
• Simulated scenarios exhibiting human-vehicle AI-based interactions and ethical dilemmas, advanced AI models for CCAM situation awareness and VRUs behavior anticipation, and car trajectory prediction.

The ultimate impact goal of AI4CCAM is to foster the adoption of Trustworthy AI in CCAM to support the European strategy for green, smart and affordable mobility by contributing to innovation and the use of data and artificial intelligence. To achieve this goal, the project will pursue five specific objectives:

• Develop an interoperable digital framework for Trustworthy AI in CCAM for managing and generating urban-traffic simulation scenarios
• Define Trustworthy AI for CCAM: Ethical, Social and Cultural Implication (ESC) and improve CAV user acceptance using biased-free datasets
• Define use cases exhibiting human-vehicle AI-based interactions and ethical issues with impact assessment
• Design trustworthy-by-design AI models for CCAM situation recognition and robust action forecast
• Boost outreach and dissemination activities, and set the basis for further developments, sustainability, and replication of the project results

3 — Concept and Methodology

In order to fulfil its ambitions, AI4CCAM is driven by a concept and methodology which aims to address the following high-level challenges:

• Challenge related to urban-traffic scenarios management and generation. The interaction between Vulnerable Road Users (VRUs) and automated vehicles is not yet sufficiently explored, especially for the Plan and Act operations of the Sense-Plan-Act paradigm. Indeed, i) existing scenarios are currently limited to the interaction with a single VRU while realistic scenarios shall involve multiple VRUs, ii) urban traffic scenarios do not involve AI-based dialogue interaction between the automated vehicle and VRUs, iii) the generation of systematic (synthetic) AI-based scenarios involving ethical dilemmas and social impact is currently insufficient.
• Challenge related to Machine Learning model robustness. Even though the usage of Deep Learning is dominant in the Sense part, the deployment of AI in real-time Plan-Act remains difficult and under-explored because of lack of robustness, explainability, safety and security. Also, used datasets for model training are verified for checking the absence of biases and augmented to improve the robustness of models.
• Challenge related to the design of trustworthy-by-design AI decision-making and action explanation. Current DL models, trained with multiple advanced datasets, are opaque and embed only little explanation means. The level of user interpretation and understanding of decisions taken by CAVs (data scientists, car manufacturers, passengers, VRUs, road regulators) is still too low. Also, ethical dilemmas and AI risks are not yet sufficiently tested in simulations.
• Challenge related to user acceptance of CAVs. User acceptance of AI-based behaviour anticipation and car trajectory prediction is not sufficiently explored, especially when multiple VRUs are involved. AI cyber-threats is also a concern for many informed road users and authorities. SSH studies are required to better understand how to remedy such challenge.